Re: [PATCH] mm: vmscan: remove deadlock due to throttling failing to make progress

[Vlastimil Babka]

On 2/3/22 19:53, Vlastimil Babka wrote:
> On 2/3/22 11:03, Mel Gorman wrote:
>> A soft lockup bug in kcompactd was reported in a private bugzilla with
>> the following visible in dmesg;
>> 
>> [15980.045209][   C33] watchdog: BUG: soft lockup - CPU#33 stuck for 26s! [kcompactd0:479]
>> [16008.044989][   C33] watchdog: BUG: soft lockup - CPU#33 stuck for 52s! [kcompactd0:479]
>> [16036.044768][   C33] watchdog: BUG: soft lockup - CPU#33 stuck for 78s! [kcompactd0:479]
>> [16064.044548][   C33] watchdog: BUG: soft lockup - CPU#33 stuck for 104s! [kcompactd0:479]
>> 
>> The machine had 256G of RAM with no swap and an earlier failed allocation
>> indicated that node 0 where kcompactd was run was potentially
>> unreclaimable;
>> 
>> Node 0 active_anon:29355112kB inactive_anon:2913528kB active_file:0kB
>>   inactive_file:0kB unevictable:64kB isolated(anon):0kB isolated(file):0kB
>>   mapped:8kB dirty:0kB writeback:0kB shmem:26780kB shmem_thp:
>>   0kB shmem_pmdmapped: 0kB anon_thp: 23480320kB writeback_tmp:0kB
>>   kernel_stack:2272kB pagetables:24500kB all_unreclaimable? yes
>> 
>> Vlastimil Babka investigated a crash dump and found that a task migrating pages
>> was trying to drain PCP lists;
>> 
>> PID: 52922  TASK: ffff969f820e5000  CPU: 19  COMMAND: "kworker/u128:3"
>>  #0 [ffffaf4e4f4c3848] __schedule at ffffffffb840116d
>>  #1 [ffffaf4e4f4c3908] schedule at ffffffffb8401e81
>>  #2 [ffffaf4e4f4c3918] schedule_timeout at ffffffffb84066e8
>>  #3 [ffffaf4e4f4c3990] wait_for_completion at ffffffffb8403072
>>  #4 [ffffaf4e4f4c39d0] __flush_work at ffffffffb7ac3e4d
>>  #5 [ffffaf4e4f4c3a48] __drain_all_pages at ffffffffb7cb707c
>>  #6 [ffffaf4e4f4c3a80] __alloc_pages_slowpath.constprop.114 at ffffffffb7cbd9dd
>>  #7 [ffffaf4e4f4c3b60] __alloc_pages at ffffffffb7cbe4f5
>>  #8 [ffffaf4e4f4c3bc0] alloc_migration_target at ffffffffb7cf329c
>>  #9 [ffffaf4e4f4c3bf0] migrate_pages at ffffffffb7cf6d15
>>  10 [ffffaf4e4f4c3cb0] migrate_to_node at ffffffffb7cdb5aa
>>  11 [ffffaf4e4f4c3da8] do_migrate_pages at ffffffffb7cdcf26
>>  12 [ffffaf4e4f4c3e88] cpuset_migrate_mm_workfn at ffffffffb7b859d2
>>  13 [ffffaf4e4f4c3e98] process_one_work at ffffffffb7ac45f3
>>  14 [ffffaf4e4f4c3ed8] worker_thread at ffffffffb7ac47fd
>>  15 [ffffaf4e4f4c3f10] kthread at ffffffffb7acbdc6
>>  16 [ffffaf4e4f4c3f50] ret_from_fork at ffffffffb7a047e2
>> 
>> The root of the problem is that kcompact0 is not rescheduling on a CPU
>> while a task that has isolated a large number of the pages from the
>> LRU is waiting on kcompact0 to reschedule so the pages can be released.
>> While shrink_inactive_list() only loops once around too_many_isolated,
>> reclaim can continue without rescheduling if sc->skipped_deactivate ==
>> 1 which could happen if there was no file LRU and the inactive anon list
>> was not low.
>> 
>> Debugged-by: Vlastimil Babka <vbabka@xxxxxxx>
>> Signed-off-by: Mel Gorman <mgorman@xxxxxxx>
> 
> Fixes: d818fca1cac3 ("mm/vmscan: throttle reclaim and compaction when too
> may pages are isolated")

And Cc: stable for 5.16.y sake

> Reviewed-by: Vlastimil Babka <vbabka@xxxxxxx>